float8: remove unneeded kernel for scale generation

Summary:

The code to create a float8 scale is unnecessarily creating an extra GPU
kernel launch by calling `torch.empty`, removing this.

Test Plan:

```
// extract trace of a linear fwd+bwd with
python benchmarks/float8/profile_linear_float8.py ~/local/tmp/test
// verify that the GPU kernel creating an empty scale tensor is no longer there

// unit tests pass
./test/float8/test_everything.sh
```

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: 3dd0c3bbf08e9e03321599b68334cf6dcc88f77b
ghstack-comment-id: 2272205849
Pull Request resolved: #616

torchao/float8/float8_utils.py

@@ -42,7 +42,6 @@ def amax_to_scale(
         float8_dtype: The float8 dtype.
         orig_dtype: The original dtype of the tensor.
     """
-    scale = torch.empty_like(amax, dtype=torch.float32)
     if float8_dtype in FP8_TYPES:
         res = torch.finfo(float8_dtype).max / torch.clamp(amax, min=EPS)
     else:
@@ -53,8 +52,7 @@ def amax_to_scale(
     # to care about this for float32/bfloat16.
     if orig_dtype is torch.float16:
         res = torch.clamp(res, max=torch.finfo(torch.float16).max)
-    scale.copy_(res)
-    return scale
+    return res.to(torch.float32)
 
 
 @torch.no_grad()